Ijiert image texture classification surf with svm

NOVATEUR PUBLICATIONS INTERNATIONAL JOURNAL OF INNOVATIONS IN ENGINEERING RESEARCH AND TECHNOLOGY [IJIERT] ISSN: 2394-3696 VOLUME 4, ISSUE 5, May-2017

IMAGE TEXTURE CLASSIFICATION: SURF WITH SVM MISS. NAMRATA N. RADE Department of Electronics and Telecommunication Engineering, Bharati Vidyapeeth’s College of Engineering, Kolhapur, India, * namratarade@gmail.com PROF. J. K. PATIL Department of Electronics and Telecommunication Engineering, Bharati Vidyapeeth’s College of Engineering, Kolhapur, India, *jaymala.p@rediffmail.com ABSTRACT Nowadays, various approaches of texture classification have been developed which works on acquired image features and separate them into different classes by using a specific classifier. This paper gives a state-of-the-art texture classification technique called Speeded up Robust Features (SURF) with SVM (Support Vector Machine) classifier. In this concept, image data representation is accomplished by capturing features in the form of key-points. SURF uses determinant of Hessian matrix to achieve point of interests on which description and classification is carried out. This method gives superior performance over already established methods in terms of processing time, accuracy and robustness. In this paper, we have taken UMD dataset for processing and calculated different performance parameters which gives excellent results.

KEYWORDS: Classification, Description, Extraction, Key-points, Matching, SURF INTRODUCTION Texture classification is a way of grouping similar things together according to common characteristics which enables us to acquire information about the image. This information can be obtained by extracting image features. With the help of features we can describe huge amount of data accurately. For acquisition of image features, texture plays a crucial role as it describes the appearance of an object. We cannot classify overlapped images separately, but they can be characterized by using their textures because each image has a specific texture that defines characteristics of an image. Texture classification has wide variety of applications in rock classification, wood species recognition, face recognition, geographical landscape segmentation, object detection and many other image processing applications [1], [2], [3], [4].

Advances in digital technology have created huge collection of digital images which requires an efficient and intelligent technique of texture classification. Texture classification is the process to divide texture features into texture classes. First stage of classification is feature extraction process in which features are extracted from image by using its texture. In second stage, features are converted into texture classes by using a classifier. Many texture classification methods have been introduced like Local Binary Pattern (LBP) [5], gray level co-occurrence matrices (GLCM) [6], Gabor filters [7], wavelet transform methods, and Independent component analysis. All these methods are based on simple computations and take more time for execution. Surf method is superior to all these earlier methods due to having advantages like robustness, computation speed and accuracy. Processing of SURF method is much faster than SIFT (Scale Invariant Feature Transform) method. SURF is an upgraded version of SIFT [8], [9], [10].

SURF is a texture detector and descriptor method which has application like object recognition, image registration, classification, reconstruction of 3D scenes and tracking objects. SURF technique handles blurred and rotational images efficiently. First process in SURF classification is integral image computation which is used to increase the performance speed. Further to extract features from an image, points of interests are achieved by using Hessian blob detector and description is obtained for each point of interest. The descriptor is based on sum of Haar wavelet responses around the point of interest. Last stage is matching of descriptors obtained from different images [11].

SURF overcomes these problems and speed up the computation with the help of integral images. Hessian matrix determinant is used in SURF to detect the location of interest points and gives stable performance as compared to other detectors like Harris detector. SURF features are extracted in terms of key-points which are found by describing the intensity distribution of pixels [12].

43 | P a g e

NOVATEUR PUBLICATIONS INTERNATIONAL JOURNAL OF INNOVATIONS IN ENGINEERING RESEARCH AND TECHNOLOGY [IJIERT] ISSN: 2394-3696 VOLUME 4, ISSUE 5, May-2017

PROPOSED METHODOLOGY In this section we present our approach. First input image is acquired from UMD dataset and then SURF features are extracted from it by using determinant of Hessian Matrix. Finally, obtained features are given to SVM classifier which converts these features into different texture classes. Simplified block diagram of this method is given below: BLOCK DIAGRAM Input Image

Extract SURF features

SVM Classifier

Fig.1: Block Diagram of proposed system PROCEDURE SURF feature classification algorithm is explained into following parts: A. IMAGE ACQUISITION: Input image is taken from UMD dataset for processing. This dataset consists of 25 different texture classes with 40 images in each class. The size of each image is 1280x960. All images from the dataset are taken with different viewpoint changes and scale differences. The dataset contains images of fruits, various plants, shelves of bottles and buckets and floor textures.

B. SURF FEATURE EXTRACTION: SURF feature algorithm consists of following parts1) Integral image computationTexture features can be promptly computed by using intermediate representation called integral image. It is very important to boost the performance speed of SURF method. Each input image is first converted into integral image and then passed for further processing. If input image and a point with coordinates x and y are given then integral image is calculated by the sum of the values between the point and the origin.

Calculation of integral image is shown in following figures:

L1

L1

L2 A

ii(x, y)

L3

L4

(i)

(ii) Fig. 2: Integral Image Calculation [(i) Integral image. (ii) Calculation of region A= L4+ L1- (L2+ L3)] If i(x, y) is an input image then integral image is given by ii(x, y) by using following formula: đ?‘–đ?‘– đ?‘Ľ, đ?‘Ś =

� ′ ≤�,�′≤�

�(� ′ , �′)....... (1)

2) Interest point detectionHessian blob detector is used by SURF to detect interest points. We have used determinant of Hessian Matrix to measure the local change around the point of interest and points are chosen with maximum value of the determinant. If we have given a point P(x,y) in an image I then Hessian matrix H(Ď ,Ďƒ) is given by, Lxx(Ď , Ďƒ) Lxy(Ď , Ďƒ) H (Ď , Ďƒ) = ....... (2) Lyx(Ď , Ďƒ) đ??żđ?‘Śđ?‘Ś(đ?œŒ, đ?œŽ) 44 | P a g e

NOVATEUR PUBLICATIONS INTERNATIONAL JOURNAL OF INNOVATIONS IN ENGINEERING RESEARCH AND TECHNOLOGY [IJIERT] ISSN: 2394-3696 VOLUME 4, ISSUE 5, May-2017

Where Lxx (ρ, σ), Lxy (ρ, σ), Lyx (ρ, σ) and Lyy (ρ, σ) are second order derivatives of gray scale image.

3) DescriptionDescriptor is used to provide a unique and robust description of an image features by describing the intensity distribution of pixels within the point of interests. This feature descriptor is depend upon Haar wavelet responses which are computed in X and Y direction around the point of interest. Description is obtained for each and every point of interest identified previously. SURF descriptors have been used widely to detect and recognize objects, faces and to extract point of interests.

4) MatchingIn the matching stage, matching pairs are found by comparing the descriptors obtained from different images.

We compare features if they are of same contrast.

C. CLASSIFICATION: Final stage of SURF algorithm is classification. Here, we have used multi-level SVM classifier which converts image features into texture classes. It compares testing set with the training set and gives proper classification of objects. SVM divides two classes by using a hyperplane. Accuracy of SVM is better than that of ANN and Nearest Neighbour Classifier.

RESULT AND DISCUSSION To analyze the performance of the proposed method, we have conducted experimets on UMD dataset.

Step1: We have implemented SURF method by using GUI. After running the code we get following GUI window. Here button ‘Init DB’ is used to start processing of whole dataset.

Fig. 3: GUI window Step 2: Take input image from dataset by pressing button ‘Test Input’

Fig.4: Test input image 45 | P a g e

NOVATEUR PUBLICATIONS INTERNATIONAL JOURNAL OF INNOVATIONS IN ENGINEERING RESEARCH AND TECHNOLOGY [IJIERT] ISSN: 2394-3696 VOLUME 4, ISSUE 5, May-2017

Step 3: SURF features are extracted in the form of key-points by using button ‘SURF’.

Fig.5: Feature Extraction Step 4: Final stage is classification of input image by using SVM classifier.

Fig.6: Classification CONFUSION MATRIX: It is a table which is used to describe the performance of a classifier by using test data. We have computed confusion matrix for five images from each class.

Fig. 7: Confusion Matrix ACCURACY OF SVM CLASSIFIER: Accuracy gives the percentage of correct classification of classifier which is the ratio of sum of true positive and true negative values to the total number of images.

Accurcy =

(đ?‘‡đ?‘ƒ+đ?‘‡đ?‘ ) đ?‘‡đ?‘‚đ?‘‡đ??´đ??ż

*100 =

43+0 50

*100 = 86%

Where, 46 | P a g e

NOVATEUR PUBLICATIONS INTERNATIONAL JOURNAL OF INNOVATIONS IN ENGINEERING RESEARCH AND TECHNOLOGY [IJIERT] ISSN: 2394-3696 VOLUME 4, ISSUE 5, May-2017

TP = True Positive= the number of samples correctly marked as positive.

TN= True Negative= the number of samples correctly marked as negative

CONCLUSION We have proposed SURF image classification techniques in which SURF feature are extracted by obtaining key-points and description is given for each detected point of interest. We have also performed classification using SVM classifier that gives accuracy of 86%. From results; we can say that SURF classification method gives best results in terms of processing speed and robustness.

REFERENCE 1) R. M. Haralick, K. Shanmugan, & L. Dinstein, “Texture Features for Image Classification,” IEEE TSMC, vol. 3, pp. 610-621, 1973.

2) Jing Yi Tou, Yong Haur Tay, Phooi Yee Lau, “Recent Trends in Texture Classification: A Review,” Symposium on Progress in Information & Communication Technology 2009.

3) T. Ahonen, A. Hadid, and M. Pietikalinen, “Face description with local binary patterns: Application to face recognition,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 28, no. 12, pp. 2037-2041, Dec. 2006 4) J. Trefny and J. Matas, “Extended set of local binary pattern for rapid object detection,” in Proc.

Comput., Vis. Winter Workshop, 2010, pp. 1-7.

5) A. Vupputuri and S. Meher, “Facial Expression Recognition using Local Binary Patterns and KullubackLeibler Divergence,” IEEE ICCSP conference, 2015.

6) Y. B. Salim and S. Nasri, "Texture classification of Woven Fabric Based on a GLCM Method and using Multiclass Support Vector Machine," SSD, 2009 7) J. A. R. Recio, L. A. R. Fernandez and A. Fernandez-Sarria, "Use of Gabor Filters for Texture Classification of Digital Images," Fisica de la Tierra, no. 17, pp. 47-56, 2005.

8) V. Gupta and K. Cecil, "An Analytical Study of SIFT and SURF in Image Registration," IJEIT, Vol.3, Issue 9, March 2014.

9) Mikolajczyk, K., and Schmid c., "An Affine Invariant Interest Point Detector", In EuropeanConference on Computer Vision(ECCV), Copenhagen, Denmark, pp. 128-142, 2002.

10) Lowe. D, "Distinctive Image Features from Scale-invariant Keypoints," IJCV 60(2004) 91-110.

11) Bay. H., Tuytelaars, T., & Van Gool, L.(2006), "SURF: Speeded Up Robust Features", 9th European Conference on Computer Vision.

12) Parul Prashar and H. Kundra, “Hybrid Approach for Image Classification using SVM Classifier and SURF Descriptor”, International Journal of Computer Science and Information Technology, Vol. 6(1), 2015, 249-251.

47 | P a g e